Machine-learned Laplacian-level density functional from exact exchange-correlation potentials and energies

Abstract

We present NNLap, a machine-learned Laplacian-level exchange-correlation (XC) functional that augments PBE with a neural-network correction depending on the electron density, its gradient, and its Laplacian. The model is trained on exact XC potentials and energies, obtained through inverse density-functional theory (DFT) calculations on configuration-interaction densities. Despite training on only a few systems -- five atoms and three molecules -- the model achieves remarkable accuracy on thermochemistry benchmarks, competing with the meta-GGA functionals SCAN and r2SCAN. It also attains accurate total energies, comparable to SCAN and better than r2SCAN and B3LYP. This shows that a Laplacian-level model, trained on exact XC potentials and energies, can reach the accuracy of meta-GGAs without their orbital dependence.